Road material cost measuring and calculating method based on generative adversarial network and conditional regression

Abstract

The invention relates to a road material cost measurement and calculation method based on a generative adversarial network and conditional regression, and belongs to the field of road material cost data measurement and calculation data processing and analysis. The method comprises the following steps: collecting road material cost data, and constructing a road surface cost data set; preprocessing the pavement cost data set to obtain engineering feature data, environment feature data and real road cost; constructing a road material cost measuring and calculating model; inputting the preprocessed data into a model, and training the model; the method comprises the following steps: calculating an error between real road cost and prediction data, measuring model performance by adopting a loss function, and optimizing parameters of each module by using an optimizer to obtain a trained model; and preprocessing to-be-predicted road data, inputting the preprocessed to-be-predicted road data into the trained model, and outputting a material cost prediction result of the road project. According to the method, high-precision prediction can be kept in a changeable engineering scene, and the generalization ability and the environmental adaptability of the model are enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of road material cost data estimation and data processing analysis, and particularly relates to a road material cost estimation method based on a generative adversarial network and conditional regression. BACKGROUND

[0002] In road construction projects, material costs usually account for a major proportion of the total project cost, so accurate estimation of material costs is crucial for project budgeting, cost control, and investment decision-making. Traditional material cost estimation methods mostly rely on manual calculation, historical data, and expert experience, which are effective in simple cases, but as engineering projects become increasingly complex, traditional methods gradually reveal their limitations, especially when faced with variable design parameters, geographical environments, and construction conditions, the accuracy and reliability of the prediction are limited. The current common road material cost estimation methods, such as experience-based estimation, statistical regression analysis, and rule-based parameterization models, while providing references in certain situations, have low accuracy and poor adaptability: experience-based methods are difficult to meet complex demands; statistical regression relies on a large amount of historical data and is not good at handling nonlinear relationships; rule-based models perform poorly under different design and environmental conditions.

[0003] In recent years, generative adversarial networks (GAN) and conditional regression models have been successfully applied in various fields, especially in generating data and making conditional predictions. GAN can generate high-quality sample data in unsupervised learning through the adversarial training of generators and discriminators; conditional regression models can accurately predict based on input condition information. In material cost estimation, GAN can generate diverse cost samples, and conditional regression models can accurately predict actual costs based on different design and environmental conditions. However, existing generative adversarial networks and regression models are mostly applied independently, and there is a lack of comprehensive solutions for the specific task of road material cost estimation. Existing technologies cannot effectively improve prediction accuracy in cases of insufficient or unevenly distributed data, and it is difficult to fully capture the nonlinear relationships between complex design and environmental conditions. Therefore, the road material cost estimation method based on generative adversarial networks and conditional regression can overcome these shortcomings and provide more accurate and reliable prediction results. By simulating cost fluctuations under different conditions through conditional generative adversarial networks (cGAN) and combining conditional regression models for accurate prediction, the stability and prediction ability of the model can be significantly improved. SUMMARY

[0004] To achieve the above-mentioned purposes, the following technical solutions are implemented: The application provides a road material cost estimation method based on a generative adversarial network and conditional regression, comprising the following steps: Collect road material cost data, construct road surface cost dataset; data preprocessing is carried out on the road surface cost dataset to obtain engineering feature data , environmental feature data and real road cost ; in the road surface cost dataset, the original engineering feature data includes road type, road grade, road width, engineering scale, road structure level and thickness, design load and material proportion; the original environmental feature data includes annual average precipitation, temperature range and humidity; the original real road cost includes concrete cost, asphalt cost, steel cost and gravel cost A road material cost estimation model is constructed, including a condition-driven data constructor, a condition-guided generation engine, an adversarial regulation joint optimizer, a multi-modal adversarial fusion device, an environmental feature optimization module and a cost-aware regression network; the engineering feature data , environmental feature data and real road cost are input into the road material cost estimation model, and the model is trained By calculating the error between the real road cost and the predicted data O, the loss function is used to measure the model performance, and the Adam optimizer is used to perform back propagation and iterative optimization on the parameters of each module until the model converges, obtaining the trained model After preprocessing the road data to be predicted, input the road data to the trained model, and output the material cost prediction result of the road project

[0005] Further, the data preprocessing operation is: encoding all non-numerical classification variables in the data set to obtain engineering feature data , environmental feature data and real road cost ; the encoding method includes label encoding and one-hot encoding

[0006] Further, the condition-driven data constructor includes a feature embedding layer, a first full connection layer, a second full connection layer, a third full connection layer and a fourth full connection layer The engineering feature data is divided into categorical features and continuous features; the categorical features include road type, road grade and design load; the continuous features include road width, engineering scale, road structure thickness and material proportion

[0007] Each feature in the categorical feature is represented by an embedding function, which maps each category index to an embedding vector; the continuous feature is uniformly projected by a linear layer to obtain a projection vector ; in the feature embedding layer, the embedding vector corresponding to each category is spliced to obtain a category vector ; the projection vector is processed by a ReLU activation function to obtain an activation vector ; the category vector and the activation vector are spliced to obtain an intermediate vector ; the intermediate vector is processed by a first fully connected layer, a second fully connected layer, a third fully connected layer and a fourth fully connected layer to obtain engineering driving features .

[0008] Further, a conditional guide generation engine and an adversarial regulation joint optimizer are adopted to effectively capture complex nonlinear feature interactions and optimize feature representation; the conditional guide generation engine includes a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer and a residual reinforcement output module. The channel attention extraction module is used to perform channel-level feature weighting on the input engineering driving features, highlight key dimensions, and suppress invalid features. The channel attention extraction module includes a first compression layer and a second expansion layer; the engineering driving features are sequentially processed by the first compression layer and the second expansion layer to obtain an expanded feature ; the expanded feature is element-wise multiplied with the engineering driving features to obtain a weighted feature ; The weighted feature is input into the first bottleneck mapping layer to obtain a first bottleneck mapping feature ; the first bottleneck mapping feature is input into the second bottleneck mapping layer to obtain a second bottleneck mapping feature ; the second bottleneck mapping feature and the engineering driving features are weighted and residual fused by the residual reinforcement output module to obtain engineering guide features .

[0009] Further, the adversarial regulation joint optimizer includes a dual-flow feature alignment module, a feature interaction fusion unit and a difference enhancement module. The dual-flow feature alignment module is used to structure the engineering driving features and the engineering guide features to make them fusible. The dual-flow feature alignment module includes a first branch path and a second branch path; the engineering driving features are processed by the first branch path to obtain a first branch feature ; the engineering driving features The first branch feature is added to obtain a first combined feature ; the second branch path includes a LayerNorm layer and a fifth fully connected layer; the engineering guide feature is normalized by the LayerNorm layer to obtain a normalized feature ; the normalized feature passes through the fifth fully connected layer to obtain a first activation feature ; The first combined feature and the first activation feature are spliced by the feature interaction fusion unit to obtain a joint feature ; the joint feature passes through the sixth fully connected layer to obtain a second activation feature ; The difference enhancement module includes a ReLU activation function and a Sigmoid activation function; the second activation feature passes through the ReLU activation function to retain feature expression and obtain a third activation feature ; the second activation feature passes through the Sigmoid activation function to generate a gating weight and obtain a fourth activation feature ; the third activation feature and the fourth activation feature are multiplied element by element to obtain an engineering optimization feature .

[0010] Further, the engineering guide feature , the engineering optimization feature and the environmental feature data are input into a multi-modal adversarial fusioner to obtain a fusion feature ; the multi-modal adversarial fusioner is introduced to optimize the fusion between different feature sources through adversarial learning, thereby improving the stability and accuracy of the model; the data processing process of the multi-modal adversarial fusioner is as follows: , wherein, represents the fusion feature; represents the L2 norm; represents a constant; · represents element-level multiplication; represents the transpose of the engineering guide feature . The multi-modal adversarial fusioner fuses information from different modalities in an antagonistic manner, enabling the model to effectively learn and optimize the relationship between multiple sources of information; by introducing an adversarial learning mechanism, the fusion feature can reflect the nonlinear interaction between various feature sources.

[0011] Further, the fusion features and the environmental feature data are input into an environmental feature optimization module to obtain environmental optimization features ; the environmental feature optimization module optimizes the environmental features, so that the model can more accurately predict under complex environmental changes; and the module adjusts the weights of the environmental features to ensure that the model can adaptively handle the influence of different environments on road material costs.

[0012] In actual engineering, environmental factors (such as geographical location, climate, etc.) have a great influence on material costs, but traditional models often ignore the changes in environmental factors. The environmental feature optimization module solves this problem, dynamically adjusts the weights of the environmental features, reduces the uncertainty of environmental factors, and improves the prediction accuracy of the model under different environmental conditions. The data processing process of the environmental feature optimization module is as follows: , wherein, represents the environmental optimization features; exp represents the exponential function.

[0013] Further, the environmental optimization features , the engineering optimization features and the engineering feature data are input into a cost-aware regression network to obtain predicted data O of road construction cost; the cost-aware regression network includes a feature fusion input layer, a deep regression modeling layer and a multi-output mapping layer; the deep regression modeling layer includes three perception networks; and the multi-output mapping layer includes a material single-item prediction head and a total cost prediction head. The environmental optimization features and the engineering optimization features are element-wise added through the feature fusion input layer to obtain first fusion features ; the first fusion features and the engineering feature data are spliced through a SiLU activation function to obtain spliced features ; the spliced features are processed through the deep regression modeling layer to obtain perception features ; in the multi-output mapping layer, the perception features are obtained through the material single-item prediction head to obtain material single-item cost prediction results; and the perception features are obtained through the total cost prediction head to obtain total cost prediction results; and the material single-item cost prediction results and the total cost prediction results are combined to obtain the predicted data O of road construction cost.

[0014] The advantages of the present application are: The present application significantly improves the accuracy and stability of road material cost prediction through the synergistic effect of various modules. Through the multi-output regression network, the model can not only accurately predict the cost of various materials, but also estimate the total cost at the same time, providing more detailed cost analysis; through the generative adversarial optimization and adversarial learning mechanism, the model's ability to capture complex nonlinear relationships is enhanced, and its robustness in the case of insufficient or unbalanced data is improved; the environmental feature optimization module enables the model to adapt to different environmental conditions, ensuring high-precision prediction in variable actual engineering scenarios, thereby effectively solving the problems of data inadaptation, overfitting and other problems in traditional methods, and enhancing the model's generalization ability and environmental adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation on the present application.

[0016] Figure 1 A step flowchart of the method of the present application; Figure 2 SHAP value contribution of different models to input features. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Embodiment 1 In this embodiment, as shown in the table, the present application provides a road material cost estimation method based on generative adversarial network and conditional regression, and the specific steps include: Figure 1 Step one, collect road material cost data, construct pavement cost dataset Rocost; data preprocessing is performed on the pavement cost dataset to obtain engineering feature data , environmental feature data and real road cost ; ​In particular, it is necessary to define the engineering feature data, environmental feature data and real road cost in detail and collect relevant data first. The original engineering feature data mainly includes relevant data reflecting the structure, scale, design and construction requirements of the road construction project. Specific parameters include: road type: expressway, urban road, rural road, branch road, bridge, road grade: national, provincial and county road, road width: (unit: meter), engineering scale: (unit: kilometer), pavement structure level and thickness: base layer, surface layer, cushion layer (unit: meter), design load: heavy load, special heavy load, material proportion: asphalt, concrete, steel bar, sandstone (unit: %). The original environmental feature data mainly describes factors related to environment and climate, which have an impact on material cost, construction progress and road life. Specific parameters include: annual average precipitation: low, medium and high, temperature range: low, medium and high, humidity: low, medium and high; The original real road cost is the cost data reflecting the actual expenditure of the project, and the specific parameters include: concrete cost: cost per cubic meter (unit: yuan / cubic meter), asphalt cost: cost per ton (unit: yuan / ton), steel cost: cost per ton (unit: yuan / ton), sandstone cost: cost per ton (unit: yuan / ton).

[0019] All data is summarized in an Excel spreadsheet, and each item of data is classified by category. Determine the unit of each item of data and unify the units. Fill in the missing values and handle the outliers to ensure data quality and avoid inaccuracies in the prediction model due to missing or abnormal data. In order to make the numerical values of different features comparable on the same scale, Z-Score standardization method is used. Divide the data into training set and test set, usually use 80%-20% proportion, 80% for training, 20% for testing, get the road cost dataset Rocost.

[0020] In particular, all non-numerical categorical variables in the dataset are encoded to obtain engineering feature data , environmental feature data and real road cost . Encoding methods include: label encoding and one-hot encoding; Label encoding Label Encoding: suitable for category features with natural order, such as road grade (high, medium, low); One-hot encoding One-Hot Encoding: suitable for category features without order relationship, such as road type.

[0021] Step two, build a road material cost estimation model, including conditional driving data configurator, condition guided generation engine, adversarial regulation joint optimizer, multi-modal adversarial fusion, environmental feature optimization module and cost perception regression network, the engineering feature data environmental feature data and real road cost Input into the road material cost estimation model, train the model.

[0022] Specifically, the engineering feature data is processed by the condition-driven data constructor to obtain engineering-driven features ; The condition-driven data constructor includes a feature embedding layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer; the output dimension of the first fully connected layer is 64, using a ReLU activation function; the output dimension of the second fully connected layer is 128, using a ReLU activation function, and the Dropout is set to 0.3; the output dimension of the third fully connected layer is 128, using a ReLU activation function; the output dimension of the fourth fully connected layer is 64, without using an activation function. The design of the condition-driven data constructor enables each layer to gradually enhance the feature extraction, fusion, and abstraction capabilities of the model. From the initial category-type data embedding to the feature optimization through the deep fully connected layer, the model can efficiently learn the complex relationships between different features. This progressive structure effectively improves the model's ability to process complex data and ensures that when processing data under different engineering conditions, it can fully capture the complex patterns hidden in the data, improving the accuracy and stability of road material cost prediction.

[0023] The engineering feature data is divided into category-type features and continuous-type features; the category-type features include road type, road grade, and design load; the continuous-type features include road width, engineering scale, pavement structure thickness, and material proportion; Each feature in the category-type features is represented by an embedding function, mapping each category index to a 16-dimensional embedding vector ; the continuous-type features are uniformly projected through a linear layer to obtain a 64-dimensional projection vector ; in the feature embedding layer, the embedding vector corresponding to each category is spliced to obtain a 48-dimensional category vector The projection vector is processed by a ReLU activation function to obtain an activation vector ; the category vector and the activation vector are spliced and merged into a 112-dimensional intermediate vector ; The intermediate vector After processing through the first, second, third, and fourth fully connected layers, the engineering-driven features are obtained. .

[0024] Specifically, the engineering-driven features The input is fed into the conditional guidance generation engine to obtain engineering guidance features. ; The condition-guided generation engine includes a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer, and a residual enhancement output module. The channel attention extraction module is used to perform channel-level feature weighting on the input engineering-driven features, highlighting key dimensions and suppressing invalid features. The channel attention extraction module includes a first compression layer and a second expansion layer; the first compression layer is configured to compress 64-dimensional features into 16-dimensional features using a fully connected layer; the parameters are: output dimension 16, activation function ReLU activation function; the second expansion layer is configured to restore the features to 64-dimensional features using a fully connected layer; the parameters are: output dimension 64, activation function Sigmoid activation function, used to generate channel attention weights. The engineering driving features The extended feature is obtained by sequentially passing through the first compression layer and the second expansion layer. The extended features With engineering-driven features Perform element-wise multiplication to obtain the weighted features. ; The first bottleneck mapping layer is implemented using a fully connected layer with an input dimension of 64 and an output dimension of 128, and the activation function is ReLU; weighted features are then processed. The input is fed into the first bottleneck mapping layer to obtain the first bottleneck mapping features. ; The second bottleneck mapping layer is implemented using a fully connected layer with an input dimension of 128 and an output dimension of 64, and does not use an activation function; it maps the features of the first bottleneck layer. The input is fed into the second bottleneck mapping layer to obtain the second bottleneck mapping features. ; The second bottleneck mapping feature and engineering-driven features After weighted residual fusion by the residual enhancement output module, the engineering guidance features are obtained. The fusion method is as follows: Preferably, Set it to 0.6.

[0025] Specifically, the engineering-driven features and the engineering guidance features The input is input into the adversarial regulation joint optimizer, and engineering optimization features are obtained ; The adversarial regulation joint optimizer includes a double-flow feature alignment module, a feature interaction fusion unit, and a difference enhancement module. The double-flow feature alignment module is used to structure the two groups of features of the engineering driving features and the engineering guiding features, so that they have fusibility.

[0026] The double-flow feature alignment module includes a first branch path and a second branch path; the first branch path includes two fully connected layers, the input dimension of the former fully connected layer is 64, the output dimension is 64, and no activation function is used; the input dimension of the latter fully connected layer is 64, the output dimension is 64, and no activation function is used; the engineering driving features The first branch feature is obtained through the first branch path ; the engineering driving features are added to the first branch feature to obtain the first merged feature ; The second branch path includes a layer normalization LayerNorm layer and a fifth fully connected layer; the input dimension of the fifth fully connected layer is 64, the output dimension is 64, and the activation function uses a SiLU activation function; the engineering guiding features are normalized by the layer normalization LayerNorm layer to obtain normalized features ; the normalized features are input into the fifth fully connected layer to obtain the first activation features ; The first merged feature and the first activation features are spliced by the feature interaction fusion unit to obtain a joint feature with a dimension of 128 ; the joint feature is input into a sixth fully connected layer to obtain second activation features ; the input dimension of the sixth fully connected layer is 128, the output dimension is 128, and the activation function uses a SiLU activation function. The difference enhancement module includes a ReLU activation function and a Sigmoid activation function; the second activation features are input into the ReLU activation function to retain feature expression and obtain third activation features ; the second activation features are input into the Sigmoid activation function to generate gating weights and obtain fourth activation features ; the third activation features and the fourth activation features Element-wise multiplication is performed to obtain the engineering optimization feature .

[0027] Specifically, the engineering guidance feature , the engineering optimization feature and the environmental feature data are input into a multi-modal adversarial fusioner to obtain a fused feature ; The data processing process of the multi-modal adversarial fusioner is as follows: , wherein, represents the fused feature; represents the L2 norm; represents a constant; and · represents element-level multiplication; represents the transpose of the engineering guidance feature . The multi-modal adversarial fusioner enables the model to effectively learn and optimize the relationship between multi-source information by adversarially fusing information from different modalities; by introducing an adversarial learning mechanism, the fused feature can reflect the nonlinear interaction between various feature sources.

[0028] Specifically, the fused feature and the environmental feature data are input into an environmental feature optimization module to obtain an environmental optimization feature ; The environmental feature optimization module enables the model to more accurately make predictions under complex environmental changes by optimizing the environmental features; by adjusting the weights of the environmental features, the module ensures that the model can adaptively handle the impact of different environments on road material costs.

[0029] In actual engineering, environmental factors (such as geographic location, climate, etc.) have a significant impact on material costs, but traditional models often ignore changes in environmental factors. The environmental feature optimization module solves this problem by dynamically adjusting the weights of the environmental features, reducing the uncertainty of environmental factors, and improving the prediction accuracy of the model under different environmental conditions. The data processing process of the environmental feature optimization module is as follows: , wherein, represents the environmental optimization feature; and exp represents the exponential function.

[0030] Specifically, the environmental optimization feature , the engineering optimization feature and the engineering feature data are input into a cost-aware regression network to obtain predicted data O of road cost; The cost-aware regression network comprises a feature fusion input layer, a deep regression modeling layer, and a multi-output mapping layer. The environmental optimization feature And the engineering optimization feature Element-wise addition is performed through the feature fusion input layer to obtain first fusion features with a dimension of 64 ; the first fusion features and engineering feature data are spliced and then passed through a SiLU activation function to obtain spliced features . The deep regression modeling layer is a core regression modeling structure, and adopts a three-layer perception network for layer-by-layer feature abstraction; the deep regression modeling layer comprises a three-layer perception network; the first layer of the perception network has an input dimension of 88 and an output dimension of 128, adopts a ReLU activation function, and has a Dropout setting of 0.3; the second layer of the perception network has an input dimension of 128 and an output dimension of 128, adopts a ReLU activation function, and uses batch normalization (Batch Normalization) to stabilize feature distribution; the third layer of the perception network has an input dimension of 128 and an output dimension of 64, and adopts a ReLU activation function; the spliced features are sequentially processed by the deep regression modeling layer to obtain perception features . The multi-output mapping layer comprises a material single-item prediction head and a total cost prediction head. In the multi-output mapping layer, the perception features are processed by the material single-item prediction head to obtain material single-item cost prediction results; the material single-item prediction head adopts a fully connected layer, has an output dimension of 4, and represents separate prediction of the cost of each material, with each output value corresponding to the cost of one material (asphalt, concrete, steel, and sandstone). The perception features are processed by the total cost prediction head to obtain total cost prediction results; the total cost prediction head adopts a fully connected layer and has an output dimension of 1; the material single-item cost prediction results and the total cost prediction results are combined to obtain prediction data O of road construction cost.

[0031] Step three, by calculating the error between the real road construction cost and the prediction data O, a loss function is used to measure the performance of the model, and an Adam optimizer is used to perform back propagation and iterative optimization on the parameters of each module until the model converges, obtaining a trained model; Step four, after pre-processing the road data to be predicted, the pre-processed road data is input into the trained model, and the material cost prediction results of the road project are output.

[0032] Example 2 The data of 1000 actual road construction projects in a province traffic engineering database in the past 5 years are selected in the embodiment, and the data are processed and divided into a training set (800 groups) and a test set (200 groups). Through the above data, the prediction performance evaluation comparison of the method of the application and the traditional regression model is carried out; the prediction performance comparison of different models is shown in Table 1. Table 1 Comparison of prediction performance evaluation of the method of the application and the traditional regression model The above table shows the performance of five common cost prediction models under the same test data set, which are respectively: linear regression (Linear Regression), support vector regression (SVR), random forest regression (Random Forest, RF), conditioned regression, and the model proposed in the application. The comparison index adopts two standards of mean absolute percentage error (MAPE) and root mean square error (RMSE), which are respectively used to measure the prediction accuracy and error fluctuation level of the model. It can be obviously seen that the method of the application is optimal in the two indexes: the MAPE is 6.2%, which is much lower than the 17.3% of the traditional linear model, the 14.6% of the support vector regression, the 12.1% of the random forest and the 10.8% of the conditioned regression, showing strong prediction stability and self-adaptive ability; the RMSE is only 48.3 yuan / m2, which is reduced by about 36% compared with the conditioned regression (75.6 yuan / m2) and the random forest (82.5 yuan / m2), and is reduced by more than 50% compared with the linear model, proving that the method has obvious advantages in error control.

[0033] The method of the application simulates the cost sample distribution that may appear under different engineering and environmental conditions through the conditional generative adversarial network, solves the problems of insufficient and unbalanced data, and enhances the generalization ability of the model to diversified scenarios. At the same time, through the fusion with the conditioned regression network, the prediction accuracy under the condition that the input features are known is significantly improved. Compared with the traditional method which only relies on historical data regression, the method can more deeply capture the complex nonlinear relationship between the design parameters and the material cost, especially when the road type, material ratio and environmental climate change significantly, the method can still maintain stable prediction effect.

[0034] The prediction deviation distribution of the model under complex environment is shown in Table 2. Table 2 Prediction deviation of the model under complex environment The above table shows the prediction deviation distribution of different models under complex environmental conditions (such as high rainfall, high humidity, extreme temperature difference, etc.). The horizontal axis represents different model types, and the vertical axis is the deviation of the predicted value minus the actual value (unit: yuan / m²). The present application selects 200 road projects under complex environments as test samples, and calculates the deviation mean, standard deviation and extreme error value of each model. It can be observed that: the traditional conditional regression model has large fluctuation in prediction deviation under complex environment, with a deviation mean of +22.5 yuan, a standard deviation of ±42.3 yuan, and a maximum error even exceeding 121 yuan / m², reflecting its instability in dealing with environmental diversity and nonlinear effects; the model proposed in the present application has a deviation mean of only +5.3 yuan, a standard deviation of ±18.7 yuan, and a maximum error controlled within 49 yuan / m², indicating that the method significantly improves the adaptability to extreme and complex conditions.

[0035] The multi-modal adversarial fusioner introduced in the method deeply fuses the engineering guidance features, optimization features and environmental variables, effectively capturing the cross-influence between environmental variables and material costs in the modeling process. At the same time, the adversarial training mechanism helps the model to learn a wider data distribution, so that the generated prediction is more "close to reality". As shown in Figure 2 As shown, the response intensity analysis of each model to the input variables is shown, using the SHAP (SHapley Additive exPlanations) value method, which reveals the contribution of different features to the model prediction result. The figure lists the 10 most key features and their average SHAP value to the model output (material cost prediction). The Y-axis represents the average SHAP value (feature contribution), and the X-axis represents the input variables of different models. The blue column represents the traditional conditional regression model; the orange column represents the model of the present application.

[0036] As can be seen from the figure, in the traditional conditional regression model, the prediction result is limited by structural factors such as road width and engineering scale, and the response to environmental characteristics (such as temperature and humidity) is weak, indicating that the environmental impact is not fully captured in the modeling; in the SHAP figure presented by the method, environmental characteristics (annual average precipitation, humidity level, and temperature range) and engineering characteristics jointly constitute the core basis for model judgment, and their average contribution is on the same order of magnitude as road structure parameters; in particular, the SHAP weight of the input variable "material ratio" is significantly improved in the method, reflecting that the model can deeply understand the influence of different material combinations on the cost structure, rather than just fitting a linear weight. The figure further verifies that the method optimizes the environmental feature optimization module and the cost-aware regression network in the design, not only improves the prediction accuracy, but also enhances the model's interpretability and engineering controllability. Users can determine which design variables are sensitive to the cost according to the SHAP analysis results, thereby assisting in optimizing engineering design and resource allocation.

[0037] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A road material cost estimation method based on a generative adversarial network and conditional regression, characterized in that, The method comprises the following steps: Collect road material cost data, construct pavement cost dataset; data preprocessing is performed on the pavement cost dataset to obtain engineering feature data , environmental feature data , and real road cost ; The road material cost estimation model comprises a condition-driven data configurator, a condition-guided generation engine, a confrontation regulation joint optimizer, a multi-modal confrontation fusion device, an environmental feature optimization module and a cost perception regression network; the engineering feature data , the environmental feature data and the real road cost are input into the road material cost estimation model, and the model is trained; By calculating the real road cost The error between the predicted data O and the actual data is measured by a loss function to evaluate the model performance, and the parameters of each module are iteratively optimized by back propagation using an Adam optimizer until the model converges, obtaining a trained model. The preprocessed road data to be predicted is input into the trained model, and a material cost prediction result of the road item is output.

2. The method of claim 1, wherein the method is based on a generative adversarial network and conditional regression. The engineering feature data is processed by a conditional driving data configurator to obtain engineering driving features :​ The conditional driving data constructor comprises a feature embedding layer, a first full connection layer, a second full connection layer, a third full connection layer and a fourth full connection layer; The engineering feature data are classified into categorical features and continuous features; Each feature in the category type feature is represented by an embedding function, mapping each category index to an embedding vector; the continuous type feature is uniformly projected by a linear layer to obtain a projection vector ; In the feature embedding layer, the embedding vector corresponding to each category is spliced to obtain a category vector ; the projection vector is processed by a ReLU activation function to obtain an activation vector ; the category vector and the activation vector are spliced to obtain an intermediate vector ; the intermediate vector is processed by a first fully connected layer, a second fully connected layer, a third fully connected layer and a fourth fully connected layer to obtain an engineering driving feature . 3.The method of claim 2, wherein, inputting the engineered driving features into a conditional guidance generation engine to obtain engineered guidance features : The conditional guide generation engine comprises a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer and a residual reinforcement output module; the channel attention extraction module comprises a first compression layer and a second expansion layer; The engineering driving feature Passing through the first compression layer and the second expansion layer in sequence to obtain an expansion feature ; The expansion feature Element-wise multiplication is performed with the engineering driving feature to obtain a weighted feature ; The weighted feature is input into a first bottleneck mapping layer to obtain a first bottleneck mapping feature ; The first bottleneck mapping feature is input into a second bottleneck mapping layer to obtain a second bottleneck mapping feature ; The second bottleneck mapping feature and the engineering driving feature are input into a residual reinforcement output module to perform weighted residual fusion to obtain an engineering guide feature .

4. The method of claim 3, wherein the method is based on a generative adversarial network and conditional regression. inputting the engineered driving features and the engineered guiding features into an adversarial regulation co-optimizer to obtain engineered optimized features : The adversarial regulation joint optimizer comprises a double-flow feature alignment module, a feature interaction fusion unit and a difference enhancement module; the double-flow feature alignment module comprises a first branch path and a second branch path; the second branch path comprises a LayerNorm layer and a fifth full connection layer; the difference enhancement module comprises a ReLU activation function and a Sigmoid activation function; The engineering driving feature The first branch feature is obtained through the first branch path The engineering driving feature The first branch feature The first combined feature is obtained by adding the engineering driving feature The engineering guiding feature The normalized feature is obtained by normalizing the engineering guiding feature through a layer normalization LayerNorm layer The normalized feature The first activation feature is obtained through the fifth fully connected layer ; The first merged feature And the first activation feature After the feature interaction fusion unit splicing operation, the joint feature is obtained ; The joint feature After the sixth full connection layer, the second activation feature is obtained ; In the difference enhancement module, the second activation feature The feature expression is retained through the ReLU activation function, and a third activation feature is obtained ; the second activation feature The Sigmoid activation function generates a gating weight, and a fourth activation feature is obtained ; the third activation feature and the fourth activation feature are multiplied element by element to obtain an engineering optimization feature .

5. The method of claim 4, wherein the method is based on a generative adversarial network and conditional regression. the engineering guidance features , engineering optimization features and environmental feature data are input into a multi-modal adversarial fuser to obtain fused features : The data processing process of the multi-modal adversarial fusioner is as follows: ， wherein, denotes a fused feature; denotes an L2 norm; denotes a constant; • denotes element-wise multiplication; denotes an engineering guided feature transpose.

6. The method of claim 5, wherein the method is based on a generative adversarial network and conditional regression. fusing the features and the environmental feature data into an environmental feature optimization module to obtain an environmental optimized feature : The data processing process of the environmental feature optimization module is as follows: ， wherein, represents an environmental optimization feature; exp represents an exponential function.

7. The method of claim 6, wherein the method is based on a generative adversarial network and conditional regression. The environment optimization feature The engineering optimization feature And engineering feature data Input into the cost-aware regression network to obtain predicted data O of road cost: The cost perception regression network comprises a feature fusion input layer, a deep regression modeling layer and a multi-output mapping layer; The environment optimization feature And engineering optimization features Element-wise addition is performed through the feature fusion input layer to obtain first fusion features ; The first fusion features Spliced with engineering feature data After splicing, the SiLU activation function is obtained Splicing features ; The deep regression modeling layer comprises three layers of perception networks; the spliced features After processing by the deep regression modeling layer, perception features are obtained ; The multi-output mapping layer comprises a material single-item prediction head and a total cost prediction head; In the multi-output mapping layer, the perception features The material single item cost prediction result is obtained through the material single item prediction head; the perception features The total cost prediction result is obtained through the total cost prediction head; the material single item cost prediction result and the total cost prediction result are combined to obtain the prediction data O of the road construction cost.

8. The method of claim 7, wherein the method is based on a generative adversarial network and conditional regression. In the road surface cost data set, the original engineering feature data comprises road type, road grade, road width, engineering scale, road surface structure level and thickness, design load and material proportion; the original environmental feature data comprises annual average precipitation, temperature range and humidity; and the original real road cost comprises concrete cost, asphalt cost, steel cost and gravel cost. 9.The method of claim 8, wherein, The data preprocessing operation is: encoding all non-numerical classification variables in all data in the data set to obtain engineering feature data , environmental feature data , and real road cost ; the encoding mode includes: label encoding and one-hot encoding.

10. The method of claim 9, wherein the method is based on a generative adversarial network and conditional regression. The categorical features comprise road type, road grade and design load; and the continuous features comprise road width, engineering scale, road surface structure thickness and material proportion.

Images